The Problem: Cloud compute for heavy AI inference is expensive, yet millions of consumer GPUs sit totally idle most of the day.
What it does: AgentFM is a single Go binary that pools those idle machines into a decentralized grid. It lets you offload heavy AI workloads from a weak machine to capable GPUs across the network without relying on central servers.
To make this work without a centralized broker, it’s built on go-libp2p using two main algorithms:
Routing & Discovery: It uses Kademlia DHT to find peers and handle NAT hole-punching, letting standard home internet connections talk directly.
State & Telemetry: It uses a GossipSub protocol where nodes broadcast their real-time hardware capacity (VRAM/CPU load). This ensures the network dynamically routes heavy tasks to capable hardware and avoids sending jobs to nodes that would instantly OOM crash.
I'm still iterating on the sandboxing for untrusted nodes, but I would love to hear what you guys think of the networking architecture!
To make this work without a centralized broker, it’s built on go-libp2p using two main algorithms:
Routing & Discovery: It uses Kademlia DHT to find peers and handle NAT hole-punching, letting standard home internet connections talk directly.
State & Telemetry: It uses a GossipSub protocol where nodes broadcast their real-time hardware capacity (VRAM/CPU load). This ensures the network dynamically routes heavy tasks to capable hardware and avoids sending jobs to nodes that would instantly OOM crash.
I'm still iterating on the sandboxing for untrusted nodes, but I would love to hear what you guys think of the networking architecture!